Indexed text attachment
games-and-information-in-higher-categories.tex
Source for Games and Information in Higher Categories
View paper · Download original file
Lines 3439–3538 of 6559. Source text is displayed literally.
3439\emph{Observable effort.} To obtain $H$ the principal pays the constant wage with $\sqrt{w}=2$: $w=4$, for a3440profit of $16-4=12$. To obtain $L$ she pays $w=1$ for a profit of $8-1=7$.3441 3442\emph{Hidden effort.} Let $u_s,u_f$ be the agent's utility from pay after success and failure. Incentive3443compatibility for $H$ is $0.8u_s+0.2u_f-1\ge0.4u_s+0.6u_f$, that is $u_s-u_f\ge2.5$; participation is3444$0.8u_s+0.2u_f-1\ge1$. Both bind at the optimum: $u_f=0$, $u_s=2.5$, so $w_f=0$ and $w_s=6.25$. The expected3445wage is $5$, and the profit is $11$. The principal still prefers high effort, but pays an \emph{agency cost}3446of $1$: the expected wage exceeds the first-best wage by exactly the risk premium the agent requires for3447carrying the incentive scheme.3448\end{example}3449 3450With a continuum of efforts and outcomes, and when the agent's problem can be replaced by its first-order3451condition, the optimal contract satisfies3452\begin{equation}\label{eq:holmstrom}3453\frac{1}{u'(w(x))}=\lambda+\mu\,\frac{f_e(x\mid e)}{f(x\mid e)}3454\end{equation}3455for multipliers $\lambda,\mu>0$ \cite{Hol79}. Pay depends on the outcome only through the likelihood ratio3456$f_e/f$: the contract rewards outcomes not because they are good for the principal but because they are3457\emph{evidence} of effort. If the likelihood ratio is increasing in $x$ (the monotone likelihood ratio3458property), pay is increasing in $x$. The first-order approach is valid under further convexity conditions3459\cite{Rog85}, and can fail without them.3460 3461\section{Informativeness}\label{sec:informativeness}3462 3463Formula \eqref{eq:holmstrom} says that the contract should use exactly the information about effort3464contained in the outcome. This can be made into a theorem that involves no first-order conditions, and it3465is here that the Markov-category language of \cref{ch:belief} earns its place.3466 3467\begin{definition}\Scl[\cite{Bla51}; categorical form \cite{Fri20,FGPR23}]3468Let $f\colon E\to X$ and $g\colon E\to Y$ be Markov kernels out of the same object (two \emph{experiments} on $E$).3469$g$ is a \term{garbling} of $f$, written $g\preceq f$, if $g=h\circ f$ for some kernel $h\colon X\to Y$. A3470statistic $s\colon X\to V$ is \term{sufficient} for $f$ if $f$ is in turn a garbling of $s\circ f$.3471\end{definition}3472 3473So the informativeness order on experiments is the factorization preorder on the coslice under $E$ in the3474category of kernels, and a sufficient statistic is a deterministic map that loses nothing in this order.3475Blackwell's theorem says that $g\preceq f$ iff every Bayesian decision maker weakly prefers observing $f$ to3476observing $g$. In an agency problem the decision maker and the source of the uncertainty are different3477people, and the corresponding statement needs its own proof.3478 3479\begin{theorem}[Informativeness principle]\label{thm:informativeness}\Scl[\cite{Hol79,GH83}]3480\begin{enumerate}[label=(\roman*)]3481\item If the monitoring technology $g\colon E\to Y$ is a garbling of $f\colon E\to X$, then for every effort $e$ the3482 least expected wage that implements $e$ under $f$ is no greater than under $g$.3483\item In particular, if $s$ is a sufficient statistic for the pair $(x,y)$ with respect to $e$, an optimal3484 contract can be taken to depend on $(x,y)$ only through $s$; and a signal $y$ has value in addition to $x$3485 only if $x$ is not sufficient for $(x,y)$.3486\end{enumerate}3487\end{theorem}3488 3489\begin{proof}3490(i) Let $w'\colon Y\to\R$ implement $e$ under $g=h\circ f$. Under $f$, consider first the \emph{random}3491contract that, on observing $x$, draws $y\sim h(x)$ and pays $w'(y)$. For every effort it induces the same3492distribution of pay as $w'$ does under $g$, so the agent's incentives and participation are unchanged. Now3493replace the random payment at each $x$ by the sure payment $\tilde w(x)$ with3494$u(\tilde w(x))=\E_{y\sim h(x)}u(w'(y))$. The agent's expected utility from every effort is the same, so3495$\tilde w$ implements $e$; and by Jensen's inequality $\tilde w(x)\le\E_{h(x)}w'(y)$, so it costs no more.3496(ii) follows by applying (i) in both directions.3497\end{proof}3498 3499The proof is two moves: pull the contract back along the garbling, then derandomize. The first uses3500only composition of kernels. The second is the one place where risk aversion enters, and it shows why3501the agency ordering can be strictly finer than Blackwell's \cite{Kim95}: what matters is not all the3502information in the signal but the information about the agent's \emph{deviations}, the dispersion of the3503likelihood ratio.3504 3505The principle has a normative reading that is often forgotten. Any variation in pay that is not evidence3506about the agent's action is pure cost. Executive pay that moves with the stock market as a whole, when the3507market is observable and not influenced by the executive, violates it; so does punishing an agent for bad3508luck that can be verified to be luck.3509 3510\section{Linear contracts}3511 3512The second-best contract of \eqref{eq:holmstrom} is typically a complicated function, sensitive to the3513exact shape of the likelihood ratio. Real contracts are simple, often linear. Holmstr\"om and Milgrom3514\cite{HM87} gave a model in which linearity is optimal, and its closed form has become the workhorse of3515applied contract theory.3516 3517\begin{example}[The linear--exponential--normal model]\label{ex:LEN}3518Output is $x=e+\varepsilon$ with $\varepsilon\sim N(0,\sigma^2)$. The agent has constant absolute risk aversion3519$r$ and effort cost $\tfrac12ce^2$; the contract is $w=\alpha+\beta x$. The agent's certainty equivalent is3520\[ \alpha+\beta e-\tfrac12ce^2-\tfrac12r\beta^2\sigma^2, \]3521so she chooses $e=\beta/c$. The fixed part $\alpha$ only transfers surplus, so the optimal slope maximizes3522total certainty-equivalent surplus $\frac{\beta}{c}-\frac{\beta^2}{2c}-\frac12r\beta^2\sigma^2$, giving3523\[ \beta^*=\frac{1}{1+rc\sigma^2}. \]3524The slope is the agent's share of the marginal product. It is $1$---sell the firm to the agent---when she3525is risk neutral or output is noiseless, and falls towards $0$ as risk aversion, noise, or the convexity of3526the cost of effort increase.3527\end{example}3528 3529Why should attention be restricted to linear contracts? In \cite{HM87} the agent controls the drift of a3530Brownian motion continuously in time and observes its path; the stationarity of the problem makes the3531optimal contract linear in final output. A different and more recent justification is robustness: if the3532principal knows only some of the actions available to the agent and evaluates contracts by their3533worst-case performance over the rest, linear contracts are optimal, because they are exactly the contracts3534that tie the two parties' payoffs together whatever the agent does \cite{Car15}. Both arguments say that3535linearity is what remains when the fine detail of the technology $f$ cannot be relied on.3536 3537\section*{Notes}3538